Detection of Wheat Single Seed Vigor Using Hyperspectral Imaging and Spectrum Fusion Strategy

Wheat is a primary staple crop in China and is pivotal in the nation's economic development. Seeds form the foundation of all agricultural activities, with seed vigor being one of the most crucial evaluation indicators. Seeds with high vigor exhibit superior field performance and storage resilience. Thus, accurately identifying wheat seeds' vigor is paramount to China's agricultural production. Traditional seed vigor detection techniques are time-consuming, demand expertise, and can irreversibly damage the seeds. Previous attempts to detect seed vigor using hyperspectral imaging technology typically focused on batch testing of seeds, utilizing either image data or spectral data, but rarely combining both for single seed vigor detection. This study explores the potential of hyperspectral imaging technology for rapid, non-destructive detection of individual wheat seeds. A total of 210 manually aged wheat seeds (105 viable, 105 non-viable) were studied. Hyperspectral data within the seeds' 400 similar to 1 050 nm band were collected, followed by a standard germination test to ensure a one-to-one correspondence between the hyperspectral data and germination results. The dataset was divided into training, testing, and real datasets in a 4:2:1 ratio. The Competitive Adaptive Reweighted Sampling (CARS) algorithm was employed to select feature bands, resulting in 30 feature bands corresponding to seed nutrients like proteins, starch, and lipids influencing seed vigor. To identify the optimal classification model, prediction models for wheat seed vigor were established using support vector machine (SVM), k-nearestneighbor (KNN), one-dimensional convolutional neural network(1DCNN), and the improved ECA-CNN machine learning algorithms, based on both full-band and feature-band spectral data from the training and testing sets. The results indicated that models built using feature-band data outperformed those using full-band data. The ECA-CNN model, constructed with feature band data, exhibited the best performance, achieving an overall accuracy of 99.17% for the training and 80% for the testing sets. The overall method and pixel method classification strategies were compared using the real dataset to negate the influence of modeling processes on comparison strategies. The findings revealed that the pixel method surpassed the overall method in detection efficacy, with an overall accuracy of 86.67%, a precision of 92.31%, and a recall rate of 80%. This research offers theoretical support for the rapid, non-destructive detection of individual wheat seed vigor.